The Decomposed Forward Pass (DePass) was proposed by Tsinghua University and the Shanghai Artificial Intelligence Laboratory in November 2025, and the relevant research results were published in the paper "DePass: Unified Feature Attributing by Simple Decomposed Forward PassIt was selected for NeurIPS 2025.

DAVSP is a unified feature attribution framework based on a single decomposition forward pass. It decomposes each latent state into additive components, propagates these components to the remaining layers, and then obtains the precise contribution of each component to the target representation. In the post-decomposition forward pass, attention scores and MLP activations are fixed, and weighted contributions are assigned based on the decomposed components. Compared to other methods, DePass achieves more faithful attribution at different granularity levels.